Uncovering high-dimensional phase space and the application of Mixture of Experts (MoE) on building the Large CALPHAD Model (LCM)

TL;DR

This paper introduces a novel multi-dimensional approach for analyzing complex alloy systems using a Large CALPHAD Model, incorporating phase space segmentation and Mixture of Experts to improve efficiency and accuracy in high-dimensional phase diagram modeling.

Contribution

It presents a new segmentation strategy based on composition space volume and integrates MoE for high-dimensional CALPHAD modeling, advancing alloy analysis methods.

Findings

Efficient modeling of high-dimensional phase diagrams for FeCoNiTi system.

Reduced computational costs through strategic segmentation.

Accurate phase behavior prediction across diverse compositions.

Abstract

This study presents a novel approach for analyzing and establishing Large CALPHAD model (LCM) in complex alloy systems. Through the introduction of "composition space volume", a multi-dimensional metric which allows to quatitatively define alloy composition variations. Utilizing stochastic methods, the study quantifies phase space complexity through phase density, and model training costs through data density. This leads to a strategic segmentation of the entire composition space, tailored to the complexity of each segment, thereby reducing computational efforts in model training. A significant advancement is the integration of segmented models using a Mixture of Experts (MoE) approach, ensuring accurate portrayal of phase behaviors across diverse composition spaces. This technique is demonstrated in establishing a high-dimensional phase diagram for the FeCoNiTi system, highlighting its efficiency and accuracy. The study's methodologies offer a systematic and cost-effective framework for modeling complex alloy systems, marking a step forward in the field of alloy design and analysis.